Graduate Theses and Dissertations Graduate College 2008 Decision making under uncertainty in power system using Benders decomposition Yuan Li Iowa State University Follow this and additional works at: http://lib.dr.iastate.edu/etd Part of the Electrical and Computer Engineering Commons Recommended Citation Li, Yuan, "Decision making under uncertainty in power system using Benders decomposition" (2008). Graduate Theses and Dissertations. 10649. http://lib.dr.iastate.edu/etd/10649 This Dissertation is brought to you for free and open access by the Graduate College at Iowa State University Digital Repository. It has been accepted for inclusion in Graduate Theses and Dissertations by an authorized administrator of Iowa State University Digital Repository. For more information, please contact digirep@iastate.edu.
Decision making under uncertainty in power system using Benders decomposition by Yuan Li A dissertation submitted to the graduate faculty in partial fulfillment of the requirements for the degree of DOCTOR OF PHILOSOPHY Major: Electrical Engineering Program of Study Committee: James D. McCalley, Major Professor Venkataramana Ajjarapu William Q Meeker JR Sarah Ryan M Leigh Tesfatsion Daji Qiao Iowa State University Ames, Iowa 2008 Copyright c Yuan Li, 2008. All rights reserved.
ii DEDICATION I would like to dedicate this thesis to my wife Ying Wu and to my parents without whose understanding and support I would not have been able to complete this work.
iii TABLE OF CONTENTS LIST OF TABLES................................... viii LIST OF FIGURES.................................. ix ACKNOWLEDGEMENTS.............................. ABSTRACT....................................... xi xii CHAPTER 1 INTRODUCTION......................... 1 1.1 Asset Management Decision Problems....................... 3 1.2 Solution Approach.................................. 5 1.3 Decision-making under Uncertainty and Risk Index................ 6 1.3.1 Risk Index.................................. 7 1.3.2 Enhanced Risk Index............................ 8 1.4 Software System Integration............................. 9 1.5 Organization of dissertation............................. 10 CHAPTER 2 POWER SYSTEM PRINCIPLES................ 11 2.1 Introduction...................................... 11 2.2 Background and Basic Conceptions......................... 11 2.2.1 Electric Power Industry........................... 11 2.2.2 Power System Reliability.......................... 12 2.2.3 N-1 Criteria.................................. 13 2.3 Power System Analysis................................ 13 2.3.1 Power Flow.................................. 14 2.3.2 Linear Sensitivity Analysis......................... 16
iv 2.3.3 Optimal Power Flow............................. 19 CHAPTER 3 BENDERS DECOMPOSITION................. 22 3.1 Introduction of the Problem Structure....................... 22 3.2 Outline of the Methodology............................. 23 3.3 Derivation of the Benders Decomposition Principle................ 27 3.3.1 Derivation of the Optimality Problem and Cut.............. 27 3.3.2 Derivation of the Feasibility Problem and Cut............... 28 3.3.3 An Example of Benders decomposition................... 29 3.3.4 Some Special Forms of Benders Decomposition.............. 31 3.4 Advantages and Limitations of Benders Decomposition.............. 32 CHAPTER 4 RISK-BASED OPTIMAL POWER FLOW.......... 33 4.1 Decomposed Security-constrained Optimal Power Flow.............. 33 4.1.1 Introduction................................. 33 4.1.2 Problem Formulation and Solution Method................ 33 4.1.3 Illustration.................................. 35 4.1.4 Conclusion.................................. 38 4.2 Improved Corrective Security-constrained Optimal Power Flow......... 38 4.2.1 Introduction................................. 38 4.2.2 Problem Formulation and Solution Procedure............... 38 4.2.3 Illustration.................................. 40 4.2.4 Conclusion.................................. 42 4.3 Risk-based Optimal Power Flow.......................... 42 4.3.1 Introduction................................. 42 4.3.2 Modeling of Severity Function........................ 44 4.3.3 Problem Formulation and Solution Method................ 45 4.3.4 Illustration.................................. 49 4.3.5 Conclusion.................................. 53 4.4 Conclusion...................................... 53
v CHAPTER 5 RISK-BASED UNIT COMMITMENT............. 56 5.1 Problem Formulation and Solving Procedure................... 56 5.2 Illustration...................................... 60 5.3 Conclusion...................................... 61 CHAPTER 6 RISK-BASED TRANSMISSION LINE EXPANSION.... 62 6.1 Risk Index for Line Expansion........................... 62 6.2 Problem Formulation and Solution Procedure................... 63 6.2.1 DC flow model and decision variable decomposition........... 63 6.2.2 Adequacy Expansion............................. 64 6.2.3 Security Expansion.............................. 64 6.2.4 Risk expansion................................ 65 6.2.5 Solution Procedure.............................. 66 6.3 Illustration...................................... 67 6.3.1 Adequacy Expansion............................. 68 6.3.2 Security Expansion.............................. 69 6.3.3 Risk-based Expansion............................ 69 6.3.4 Sensitivity Analysis............................. 70 6.4 Approach Extensions................................. 71 6.4.1 Multiple Load Scenarios........................... 72 6.4.2 Integration of Operating Cost and Candidate Generator Capacity... 73 6.4.3 Transmission and Generation Expansion.................. 74 6.4.4 Computation Speed Analysis........................ 75 6.5 Conclusion...................................... 75 CHAPTER 7 RISK-BASED VAR RESOURCE ALLOCATION AND VAR GAP......................................... 76 7.1 Risk Index for the Var Resource Allocation.................... 76 7.2 Problem Formulation and Solution Procedure................... 79 7.2.1 Problem Formulation: Adequacy and Security Expansion........ 79
vi 7.2.2 Problem Formulation: Risk Expansion................... 80 7.2.3 Solution procedure.............................. 81 7.3 Illustration...................................... 82 7.3.1 Adequacy Expansion............................. 83 7.3.2 Security Expansion.............................. 83 7.3.3 Risk-based expansion............................ 83 7.3.4 Sensitivity analysis of the risk-based approach.............. 84 7.3.5 Comparison.................................. 85 7.4 Var Gap........................................ 85 7.4.1 The Concept of Var Gap........................... 86 7.4.2 Calculation of VVG............................. 87 7.4.3 Calculation of SVG.............................. 87 7.4.4 Different Types of Var Allocation Formulation Using VGs........ 88 7.5 Illustration of Var Allocation and VGs....................... 89 7.5.1 Type II Illustration.............................. 89 7.5.2 Type IV Illustration............................. 90 7.6 Conclusion...................................... 90 CHAPTER 8 GENERAL BENDERS DECOMPOSITION STRUCTURE AND SOA BASED PLATFORM........................ 92 8.1 General Benders Decomposition Structure..................... 92 8.1.1 Introduction................................. 92 8.1.2 GBDS Elements............................... 93 8.1.3 GBDS Information Communication.................... 96 8.1.4 Structured Programming using GBDS................... 97 8.1.5 GBDS Unbundled Services......................... 97 8.1.6 Conclusion.................................. 97 8.2 BenSOA A Decision Making Paradigm....................... 98 8.2.1 Introduction................................. 98
vii 8.2.2 Service Oriented Architecture........................ 99 8.2.3 Benders Decomposition and SOA: BenSOA................ 101 8.2.4 Service Inventory............................... 102 8.2.5 Structure Design of BenSOA........................ 104 8.2.6 BenSOA and Decision Making System................... 104 8.2.7 Conclusion.................................. 105 CHAPTER 9 CONCLUSION........................... 106 9.1 Contributions..................................... 106 9.2 Further Research Direction............................. 107 BIBLIOGRAPHY................................... 108
viii LIST OF TABLES Table 3.1 Problem Formulation Summary for BD, GBD and LGBD....... 23 Table 3.2 Estimation of the α(x)........................... 31 Table 4.1 Computing speed and results using different strategies of RTS96.... 37 Table 4.2 Results of different strategies....................... 41 Table 4.3 Circuit parameters............................. 49 Table 4.4 Generation unit information........................ 50 Table 4.5 Bus information............................... 50 Table 6.1 Generation capacity and loads....................... 68 Table 6.2 Candidate transmission line information................. 69 Table 6.3 Line failure probability........................... 71 Table 6.4 Expansion Considering Operation Cost and New Unit Capacity.... 74 Table 7.1 Candidate Var cost information ($M)................... 84 Table 7.2 Line outage probability (1E-3)....................... 84 Table 7.3 Cost comparison for different strategies.................. 85 Table 7.4 Summary of Var resource formulation................... 88 Table 7.5 Collapse point voltage effects....................... 89 Table 7.6 Cost comparison for different strategies for type IV Var allocation... 90
ix LIST OF FIGURES Figure 1.1 Illustration of risk index and decision-making.............. 7 Figure 1.2 Probability density function plot of the risk............... 8 Figure 2.1 SFT flowchart................................ 19 Figure 3.1 Benders process............................... 26 Figure 3.2 Benders procedure............................. 26 Figure 3.3 Graph description of feasibility cut.................... 28 Figure 3.4 Example of the Benders decomposition.................. 31 Figure 4.1 The flowchart to solve SCOPF parallel with Benders Decomposition. 35 Figure 4.2 Computation time comparison of 6-bus test system........... 36 Figure 4.3 Computation evolution of DSCOPF.................... 37 Figure 4.4 Flowchart to solve ICSCOPF with Benders Decomposition....... 40 Figure 4.5 Sensitivity analysis............................. 42 Figure 4.6 The system operation state........................ 44 Figure 4.7 Solution procedure of RBOPF using Benders decomposition...... 48 Figure 4.8 The system operation state cut by Benders cut............. 52 Figure 4.9 The solutions of all the cases........................ 53 Figure 4.10 Sensitivity analysis of the risk-based optimal power flow........ 54 Figure 4.11 Sensitivity analysis of the risk-based optimal power flow........ 55 Figure 5.1 Flowchart of Solving RBUC using Benders decomposition....... 59 Figure 5.2 The convergence of the three-stage RBUC................ 60
x Figure 5.3 Risk of next 24 hours for 7 contingencies................. 61 Figure 5.4 The effect of a real-time failure probability................ 61 Figure 6.1 The illustration of the overload severity function............. 63 Figure 6.2 Flowchart of the approach for line expansion............... 67 Figure 6.3 The 6-bus test system for line expansion................. 68 Figure 6.4 The effect of different line expansion strategy.............. 70 Figure 6.5 Risk sensitivity comparison........................ 71 Figure 7.1 Relation of bus voltage with loadability.................. 77 Figure 7.2 The illustration of the voltage collapse severity function........ 78 Figure 7.3 The relationship of voltage, loadability and Var injection........ 78 Figure 7.4 The flowchart of the risk-based Var resource allocation......... 82 Figure 7.5 The AEP 14-bus test system........................ 83 Figure 7.6 Voltage risk sensitivity analysis...................... 85 Figure 7.7 The effect of reactive power shown by PV curve............. 86 Figure 7.8 The collapse point at different voltage level............... 88 Figure 7.9 Type IV sensitivity analysis........................ 91 Figure 8.1 Mapping between decision triplet and Benders triplet.......... 95 Figure 8.2 Decision making using GBDS....................... 96 Figure 8.3 The hockey stick function......................... 99 Figure 8.4 Service and service design principles................... 100 Figure 8.5 BenSOA services inventory......................... 102 Figure 8.6 Structure design of BenSOA........................ 105
xi ACKNOWLEDGEMENTS I would like to take this opportunity to express my thanks to those who helped me with various aspects of conducting research and the writing of this thesis. First and foremost, Dr. James D. McCalley for his guidance, patience and support throughout this research and the writing of this thesis. His insights and words of encouragement have often inspired me and renewed my hopes for completing my graduate education. I would also like to thank my committee members for their efforts and contributions to this work: Dr. Ajjarapu Venkataramana, Dr. Meeker William Q JR, Dr. Ryan Sarah M, Dr. Leigh Tesfatsion and Dr. Qiao Daji.
xii ABSTRACT Decision-making for operations, maintenance, and investment planning of electric power systems must handle a great deal of uncertainty. In the work described here, the enhanced risk index is used to describe these uncertainties, and the Benders decomposition algorithm plays the role of integrating three components of the decision making problem: economy, reliability, and risk. A decomposed security-constrained optimal power flow is developed to demonstrate the significant speed enhancement of the chosen algorithm. The risk-based optimal power flow, risk-based unit commitment problem, risk-based transmission line expansion, and risk-based Var resource allocation are formulated and demonstrated. A general Benders decomposition structure is developed to cover most of the decision making problems encountered in everyday use within the power industry. In order to facilitate this algorithm, a service oriented architecture (SOA) is introduced and a Benders decomposition and SOA based computation platform is designed.
1 CHAPTER 1 INTRODUCTION Electric power, natural gas, oil, ground and air transportation, communications, and water reservoirs are critical national infrastructures, each of which has a large number of distributed, interdependent, capital-intensive physical assets that can fail in catastrophic ways. The common asset management challenge is a set of decision problems related to operation, maintenance, and planning of those assets where decision-makers must identify alternatives and for each one, assess costs, benefits, and risks. Quality of resulting decisions depends on quality of information used in the assessments and how that information is processed. Central, and essential, are information characterizing the health, or condition, of the assets. Often-used indicators of asset condition are age and time since the last inspection and maintenance, so nameplate data and maintenance histories have always been highly influential in the decision process. Condition data from manual inspections are also incorporated if available. It has been only recently, however, that sensing, communication, and database technology has evolved to the point where it is feasible for decision-makers to access operating histories and asset-specific real-time monitoring data. Creative use of this data via processing, fusion, assessment, and decision algorithms can significantly enhance the quality of the final actions taken and decision-maker confidence, and, for even one of the aforementioned industries, result in very large national impact in terms of more economic and reliable system performance. A rough estimate of the numbers of power transformers and circuit breakers comprising the US transmission system (138-765 kv) are 150,000 and 600,000, respectively; in addition, there are 254,000 miles of high voltage transmission lines. Total replacement value of the lines alone (excluding land) is conservatively estimated at over $100 billion dollars [1] and triples when including transformers and circuit breakers. Investment in new transmission equipment has
2 significantly declined over the past 15 years. Some of the equipment is well beyond intended life, yet is operated under increasing stress, as load growth, new generation, and economically motivated transmission flows push equipment beyond nameplate limits. Economic operation, and ultimately electric energy price, is heavily influenced by transmission equipment availability, because transmission forced outages require utilization of more expensive generation. Maintaining acceptable electric transmission system reliability and delivering electric energy at low energy prices requires innovations in sensing, diagnostics, communications, data management, processing, algorithms, risk assessment, decision-making (for operations, maintenance, and planning), and process coordination. Among these asset management problems, the data-driven electric power industry has made strides in sensing and diagnostics. Yet there has been less progress in (a) communications, (b) data management, information processing and associated algorithms, (c) risk assessment methods, and (d) decisionmaking paradigms, and progress has been almost nonexistent in (e) process coordination. This dissertation will focus on the latter three of these by developing and linking multi-timescale stochastic decision algorithms. There are three interconnected electric power transmission grids in North America: the eastern grid, the western grid, and Texas, with each being comprised of assets which include transmission lines, support structures, transformers, power plants, and protection equipment. Within each grid, power supplied must equal power consumed at any instant of time; also, power flows in any one circuit depend on the topology and conditions throughout the network. This interdependency means that should any one element fail, repercussions are seen throughout the interconnection, affecting system economic and engineering performance. Overall management requires decision in regards to how to operate, how to maintain, and how to reinforce and expand the system, with objectives being risk minimization and social welfare maximization. The three decision problems share a common dependence on equipment health or propensity to fail; in addition, their solutions heavily influence future equipment health. As a result, they are coupled, and optimality requires solution as a single problem. However, because network size (number of nodes and branches) together with number of failure states
3 is so large, such a problem, if solved using traditional optimization methods, is intractable. In addition, the three decision problems differ significantly in decision-horizon, with operational decisions implemented within minutes to a week, maintenance decisions within weeks to a couple of years, and investment decisions within 2-10 years. Therefore, excepting the common dependence and effect on equipment health, the coupling is sequential, with solution to latterstage problem depending on solution to former-stage problems. Because of this, the industry has solved them separately, with the coupling represented in a very approximate fashion via human communication mechanisms. It should be the case, then, that traditional solutions are suboptimal. The objective of this work is to design and develop a system for integrating decision algorithms associated with power system risk-reliability-economy decision problems. Doing so will improve all decisions. In accomplishing this objective, we perform a decomposition of the essential decision problems and we provide a solution algorithm for each of them. The various decision problems are then tied together using Benders decomposition, and a service orientated software architecture is designed to support these decision problems. An assumption made in this dissertation is that the failure probabilities of the components are known. It is recognized, however, that obtaining probabilities accurately reflecting equipment state for a given type of decision problem is a non-trivial problem, but it is one that is left to others in order to limit the scope of this work. The decision models developed in this work are intended to operate interactively, so that decision in each time frame utilizes information from decision within other time frames. 1.1 Asset Management Decision Problems Asset management decision problems are classified into one of 4 types which all involve resource allocation with the objective to minimize cost and risk. These specific asset management decision problems include (a) Operations, (b) Short-term maintenance selection and scheduling, and long-term maintenance planning, and (c) Facility planning. These problems differ primarily in their time scale but are linked by a common focus on the interactions be-
4 tween the condition of equipment and the decisions taken. The work of this dissertation will focus on the operations and the facility planning problems. The operations problem is to allocate supply in the next hour-to-week to meet the demand while minimizing risk (exposure to failure consequences); the decision variables are the generation levels at the power plants. There are two levels of this problem. The first is the unit commitment (UC) problem. The second is the optimal power flow (OPF) problem. OPF is actually a UC sub-problem, but UC is typically solved on a weekly basis, resulting in identification of which units are connected at each hour, with a simplified version of OPF. OPF is solved on at least an hourly basis, to find the allocation of energy among the connected units, given the UC solution. In the OPF, the objective is to minimize the generation cost of supplying a specified system power demand subject to power flow equations (based on the Kirchoff s laws governing the power flowing in the network) together with hard constraints on system and asset capabilities [2]. The security-constrained OPF (SCOPF) is the OPF with additional operational constraints for each of some identified network contingencies. The UC is an integer program (IP) to select, for each hour of a week, the units to be connected to the grid. The literature is rich in work on this problem, with [3, 4] providing good summaries. Yet little work has been done on the integration of UC and transmission reliability, with [5, 6] being exceptions that model security as hard constraints rather than as an objective to be achieved. The security-constrained UC (SCUC) is the UC with additional operational constraints for each of some identified network contingencies. The long term transmission planning problem under uncertainty, to allocate financial resources over a time horizon on the order of a decade to provide adequate transmission facilities, can be formulated as a stochastic program [7]. As in [8], it is typical to employ a DC load flow approximation to obtain a more tractable linear model. The closely related reactive power planning will be formulated as a nonlinear stochastic program.
5 1.2 Solution Approach J. F. Benders [9] proposed the Benders decomposition in 1962 and A. M. Geoffrion [10] extended it in 1972, following which it began to find applications in power system decisionrelated problems. The basic idea underlying Benders decomposition is to solve large-scale problem via a sequence of smaller ones. This provides that Benders is well-suited to integrating the various power system asset management decision algorithms. This dissertation is not the first work to utilize Benders for power system decision problems, and in fact it has up to now experienced the following three stages with respect to the power system area. I. Mathematical development stage (1960 s and early 1970 s): This stage includes the invention of Benders decomposition method in 1962 and further development in 1972. II. Early application in power systems (1980 s and 1990 s): This stage resulted in some good applications of Benders decomposition to power systems including hydrothermal coordination, corrective security-constrained optimal power flow, maintenance scheduling, generation, transmission, and Var planning are developed. III. Present application in power system (late 1990 s until now). Most works focus on SCUC during this period, and maintenance scheduling, and planning are reformulated for competitive market systems. References [11, 12] give solutions to the dispatch and scheduling of hydrothermal system and give good explanations of the Benders method. The generation planning problem is solved in [13, 14, 15, 16, 17, 18]. CSCOPF (corrective security-constrained optimal power flow) is proposed in [19]. Reference [20] is a good summary: the Benders decomposition is reviewed and the applications to CSCOPF, transmission planning, Var allocation, and parallel computation are described. Var planning is described in [21, 22, 23] and in [24, 25] voltage stability is considered. References [26, 27, 28, 29, 30] use Benders to solve the transmission planning problem. It is also used to address composite generation-transmission planning in [31, 32, 33]. It is used to address the maintenance problem in [34, 35, 36, 37, 38, 39]. ATC is calculated
6 using Benders in [40]. References [41, 42, 6, 43, 44, 45, 46, 47, 48, 49, 50] use Benders to solve the unit commitment problem considering network and security constraints. Reference [51] is another good summary of Benders applications in power systems. References [52, 53, 54, 55] use Benders to integrate risk into the OPF, SCUC, and planning. The application of Benders decomposition in this dissertation is unique, relative to that of the literature, in that the risk index [56] is integrated to address the uncertainty and that the economy, reliability, and the risk management are systematically integrated in each decision problem addressed. In addition, the use of Benders is facilitated by the service-oriented programming approach used to develop certain kinds of software systems today. 1.3 Decision-making under Uncertainty and Risk Index Operating and planning bulk interconnected electric power systems are complex activities. Conflicting issues such as economy and reliability have to both be considered when making the appropriate decisions. Although economic competition is emphasized in the restructured power market environment, reliability remains the principal core value of the power industry. How to make a particular decision in a economic manner while maintaining reliability under uncertainty remains an important power industry problem. Generally the decision variables can be placed into two categories: 1. Amount of control that should be applied for available resources; 2. Whether existing components should be maintained or replaced or whether new components should be added. Both these decision variables categories apply to economic related decisions. Both integer and continuous decision variables are involved in decision-making problems. The problems could be specified in the form of the linear programming (LP), nonlinear programming (NLP), mixed integer linear programming (MILP) and mixed integer nonlinear programming (MINLP). The objective functions of these decision-making problems are typically economic benefits such as cost, investment and social welfare. The constraints are typically
7 expressed as power flow equation, flow limits, and system parameter limits. These problems could be very complex. Decision-making without considering uncertainty is impractical in most situations especially as the time span increases. All network-related decision making in power systems must consider the unexpected failures of power system components and load fluctuations. The risk index is used in this work to describe these uncertainties. 1.3.1 Risk Index Risk is defined as the product of probability and severity [56], shown in (1.1): R = P Sev (1.1) where, R is the risk index, P is the probability with respect to future load scenarios and outages, and Sev is the severity of the future situation. As shown in Fig.1.1, decision makers have to deal with the following situation: they make a decision now and suffer the corresponding impact in the future. So pursuing the single objective of lowering cost could result in huge future risk. Alternatively, avoiding all future risk definitely will make the cost too high. The role of the risk index is therefore to serve as an indicator which biases the decision conservatively when risk is too high and optimistically when risk is comparatively low, according to the given information. Now Future Operation (u0) Planning (x0) Operaton & Planning Decision Pn P 1 Cond 1 Cond n P c P c P 1 P 1 Cont 1 Sev Cont c Sev Cont 1 Sev Cont c Sev Cost Risk Time Possible future operating cdts (bus injection) Selected future contingency states Figure 1.1: Illustration of risk index and decision-making
8 In Fig.1.1, the decision variables for now can be put into two categories as introduced before. Although the severity might also have economic consequences, they are restricted in this dissertation to power system reliability consequences in this dissertation such as over flow and load shedding. 1.3.2 Enhanced Risk Index The risk index defined in [56] is a pure expectation index that provides a very good indication of the current decision. When risk index is integrated within the optimal decision problem to facilitate the decision making, the risk can be managed through the adjustment of the candidate decisions. For example, placing a heavier weight on the risk part could lead to a more conservative and higher cost decision to reduce the risk. At the same time, in the optimization procedure both aspects of the risk index can be managed respectively. For example, the probability of component failure can be reduced by appropriate maintenance [57]. More important, in power system reliability-related decision-making, some level of management of the severity is highly needed in the decision-making process. As shown in Fig.1.2, situation C outperforms A and B in both expectation and variance, and A has a smaller expectation but larger variance. C is the most preferred situation. If a severity tolerance is set, situation B should be chosen in favor of avoiding huge severity between A and B. But if it is known that the severity can be relieved, A will be chosen. Pr μ C μ A μ B Severity tolerance Sev Figure 1.2: Probability density function plot of the risk In this dissertation, two types of severity management are provided. Controllability represents the situation that after an event, the severity can be mitigated
9 by available control resources. Acceptability means that if a particular events happens, just how much severity should be allowed to happen. Other situations may also occur. Sometimes the controllability may not be sufficient to provide complete mitigation of the severity, resulting in existence of a certain level of tolerance for the severity. After the management of severity is integrated into the risk index, it is renamed the enhanced risk index. In this dissertation, the enhanced risk index is used in almost all the cases and it will still be called simply the risk index. Under uncertainty, decision-making problems such as those described will become more complex and remains a hot topic in the power industry. In the deregulated environment, decision-making under uncertainty has become more attractive due to both economic motivation and computational development of the modern computer. This dissertation aims to provide systematic answers for these decision-making problems. 1.4 Software System Integration When applying decomposition techniques to the real practical large scale problems, how the parallel computation is performed, how these smaller problems interact with each other, how the security is ensured, how the user can interface with these problems become an essential problem. The need for a structure, a protocol, or a diagram is needed. The information technology is also evolving. The structured programming, objective oriented programming and service oriented architecture (SOA) recently and finally. The service oriented architecture provides a level of modularity and autonomy that is attractive for decision algorithms encountered in power systems. The spirit of Benders decomposition is to decompose the large and complex problem into loose coupling small and easy problems. All the decomposed problems are autonomic problems and typically independent with each other. At the same time, these subproblems are stateless
10 and composible. These characteristics are similar to the principles of the SOA. Based on these facts, a Benders decomposition and SOA based decision making diagram, BenSOA, is proposed in this dissertation. In this design, the subproblems of Benders decomposition will be the services of SOA. The SOA infrastructure will facilitate the communication between subproblems and realize the decision making process. 1.5 Organization of dissertation The rest of this dissertation is organized as follows: Chapter 2 introduces the background of power industry, basic principles of the power system engineering, and decision-making under uncertainty, Chapter 3 describes Benders decomposition algorithm, Chapter 4 discusses the decomposed security-constrained optimal power flow, the improved corrective securityconstrained optimal power flow, and the risk-based optimal power flow, Chapter 5 proposes the risk-based unit commitment, Chapter 6 addresses risk-based transmission system expansion planning, Chapter 7 proposes risk-based Var allocation, Chapter 8 proposes the general Benders decomposition structure and introduces the design of the Benders decomposition and service oriented structure integrated platform, and chapter 9 provides conclusions and suggestions for future work.
11 CHAPTER 2 POWER SYSTEM PRINCIPLES 2.1 Introduction An understanding of the structure of the power industry and the basic conceptions of power system reliability will facilitate decision making. Meanwhile the power system modeling and analysis play an important role in the decision making problems because all the decision makings in power system area must consider the reliability issue. The linear sensitivity analysis and simultaneous feasibility test (SFT) used in power system is introduced, which can be thought as a simple mimic form of Benders decomposition. In this chapter, these basic principles will be introduced and some of these will be used in the later chapters. 2.2 Background and Basic Conceptions 2.2.1 Electric Power Industry In the traditional power industry, vertically integrated electric utilities managed three main power system components: generation, transmission, and distribution within its own territorial monopoly. In the restructured power industry, these three main components are unbundled and the new entities such as the generation company, the transmission company, and the load service entity, take charge in the three components respectively and represent their own individual interests. A new entity, typically called an independent system operator (ISO), also emerges to operate the system and to maintain system reliability. An electric market is created through the basis of the unbundling of the principal components of the power industry. The market has a much larger territory than that of the traditional utility. Two kinds of economic entities are created: the players and the ISO. The
12 players include the generation company, the transmission company, the load service entity, the broker, etc. These players each represent themselves and pursue their own benefit in the electric market. The ISO represents society and tries to simultaneously maximize social welfare and maintain power system reliability. 2.2.2 Power System Reliability The NERC s definition of the electric system reliability [58] is: Reliability The degree of performance of the elements of the bulk electric system that results in electricity being delivered to customers within accepted standards and in the amount desired. Reliability may be measured by the frequency, duration, and magnitude of adverse effects on the electric supply (or service to customers). According to the NERC s definition, reliability can be addressed by considering two basic and functional aspects of the electric system adequacy and security: Adequacy The ability of the bulk electric system to supply the aggregate electrical demand and energy requirements of customers at all times, taking into account scheduled and reasonably expected unscheduled outages of system components. Security The ability of the bulk electric system to withstand sudden disturbances such as electric short circuits or unanticipated loss of system components or switching operations. In plain language, adequacy implies that sufficient generation and transmission resources are available to meet projected needs plus reserves for contingencies. Security implies that the power system will remain intact even after outages or equipment failures. One of the main aspects of system security is so-called steady-state security [59]. This is defined as the ability of the system to operate steady-state-wise within the specified limits of safety and supply quality
13 following a contingency, in the time period after the fast-acting automatic control devices have restored the system load balance, but before the slow-acting controls, e.g. transformer tappings and human decisions, have responded. This dissertation focuses on steady-state security for power systems operation and planning. 2.2.3 N-1 Criteria Generally the security represents the so-called N-1 criteria. N is the total number of transmission elements in the system and N-1 is the total system with one element out of service. The minus one could be a generation unit, a transmission line or a transmission transformer. The basic idea is that even if one component is lost, the system should still satisfy the load requirements without operating violation. Transmission line and transformer failures generally cause changes with respect to the flows and voltages on the transmission equipments still remaining connected to the system. Generation failures can also cause flows and voltages to change in the transmission system, with possible addition of dynamic problems involving system frequency and generator output. In this dissertation, the security assessment will use this criteria. 2.3 Power System Analysis Modern power systems have become increasingly complex and interconnected over the past one hundred years. These systems, however, are evolved from some very basic theories developed by some scientific giants more than one hundred years ago and these basic theories remain the unchanged bases of the power system analysis. Higher level layers have of course been added to the power systems in the past and will be in the future. Some basic modeling and analysis methods will be introduced in this section as preparation for the work discussed in the later chapters.
14 2.3.1 Power Flow In power flow analysis the transmission system is modeled as a set of buses or nodes interconnected by transmission links. Generators and loads, connected to various nodes of the system, inject and withdraw power from the transmission system. The compact form of the power flow equations is as follows: g(x) = 0 (2.1) In the equation (2.1), algebraic variables x represent the solution of power flow. Algebraic equations g represent network equations. At each bus of the power system, power injection is balanced. 2.3.1.1 AC Power Flow For AC power flow, the network equations (2.1) can be expanded into nonlinear forms representing both real power and reactive power balance. P gi P li P ti = 0 Q gi Q li Q ti = 0 (2.2) where P gi is real power generation at bus i, P li is real power load at bus i, P ti is net real power injection at bus i, Q gi is reactive power generation at bus i, Q li is reactive power load at bus i, and Q ti is net reactive power injection at bus i. The real and reactive power generations are determined by the inherent characteristics of the generator. The real and reactive loads are determined by the load characteristics. The net real and reactive power injections are constrained by the physical characteristics, which are represented by the following equations: P ti = N j=1 V iv j [G ij cos(θ i θ j ) + B ij sin(θ i θ j )] Q ti = N j=1 V iv j [G ij sin(θ i θ j ) B ij cos(θ i θ j )] (2.3) where θ i and θ j are the phase angles at buses i and j respectively, V i and V j are the bus voltage magnitudes respectively, and G ij + jb ij = Y ij is the ij term in the Y bus matrix of the power system.
15 The variables V and θ are bus voltage and angle respectively, and these variables belong to the unknown variables x in (2.1). Generally if V is given the correspondent reactive power balance equation will be removed from (2.1) and if θ is given the correspondent real power balance equation will be removed from (2.1). 2.3.1.2 DC Power Flow DC power flow, which is quite fast to analyze and therefore permits determination of a large series of power flows within a reasonable time frame, is a good approximation to AC power flow under normal load situations. The following assumptions are made for DC power flow: All branch resistances are equal to zero. All voltage magnitudes are constant. The differences of phase angles between voltages at the ends of any branch are within normal loading range (where the errors are not very high). Under these assumptions, there are no losses in the system (no resistance) and a real power solution can be obtained without solving simultaneously for reactive power. Using nomenclature of DC power flow, the B matrix of DC power flow is: B = r y r T (2.4) where r is the reduced node-branch incidence matrix. The DC power flow is formulated simply as follows (reference bus removed): P = B θ (2.5) where θ is the bus angle vector (reference bus not included), and the P is the bus injection vector (reference bus not included). The line flow is as follows: p ij = y ij (θ i θ j ) (2.6)
16 where y ij is the admittance between bus i and j. The vector form of line flow is as follows: P ij = y r T B 1 P = H B 1 P = H θ (2.7) Although the reference bus voltage angle is known and will not appeared in vector θ, all the line flow can be obtained using equation (2.7) due to the characteristics of DC flow. DC power flow is useful for rapid calculations of real power flows and is very useful in security analysis studies. 2.3.2 Linear Sensitivity Analysis There are two types of linear sensitivity analysis: linear sensitivity coefficients and linear sensitivity factors corresponding to AC power and DC power flow respectively. These sensitivities will be used in the simultaneous feasibility test process. One important observation is that the idea of the linear sensitivity analysis is very similar to Benders decomposition used in power system area. Both linear sensitivity analysis and Benders decomposition use the gradient or Lagrange information to adjust the tentative results. 2.3.2.1 Linear Sensitivity Coefficients Linear sensitivity coefficients give an indication of the change in one system quantity (e.g., MW flow, MVA flow, bus voltage, etc.) as another system quantity (e.g., generator MW output, transformer tap position, etc.) is varied. One important assumption is that as the adjustable variable is changed the power system reacts so as to keep all of the power flow equations solved [3]. One example deriving linear sensitivity coefficients of an AC network model is given here. The power flow equation is repeated as: g(x, u) = f(x) u = 0 (2.8) where x is the state vector of voltages and phase angle and u is the vector of control variables. For simplicity, the control variable here is defined as the generator MW output.
17 From the current operating point x and u, a small enough adjustment occurs and before and after power flow is solved, yielding: g(x +, u + ) = g(x, u) + g (x, u) [ x, u] T = 0 g (x, u) [ x, u] T = [f (x), 1] [ x, u] T = 0 f (x) x = u J fx x = u Transmission system dependent variables, h, are assumed, which can be MVA flows, load bus voltages, etc, and can be expressed as a function of the state and control variables. We try to find their sensitivity with respect to changes in the control variables. Around the current operating point, we have: According to the above derivation, we get h = J hx x + J hu u h = [J hx J 1 fx + J hu] u h = J u (2.9) where J is the linear sensitivity coefficients matrix between h and u around the operating point. One important characteristic of the linear sensitivity coefficients is that the value is only good for small adjustment and the sensitivities must be recalculated often [3]. 2.3.2.2 Linear Sensitivity Factors Similar to DC power flow, determination of network linear sensitivity factors is another fast approximation method used in power flow problems. There are several kinds of sensitivity factors, including generation shift factors and line outage distribution factors [3]. Only generation shift factors are described in this example. The GSF (generation shift factors) are designated α li and have the following definition: α li = f l P i (2.10)
18 where l is the line index, i is the bus index, f l is the change in megawatt power flow on line l when a change in generation, P i, occurs at bus i, and P i is the change in generation at bus i. It is assumed in this definition that the change in generation, P i, is exactly compensated by an opposite change in generation at the reference bus, and that all other generators remain fixed. The GSF at the reference bus is zero. The generation shift factors are linear estimates of the change in flow resulting from a change in power at a bus. The effects of simultaneous changes on several generating buses can be calculated using superposition. A special situation with respect to the generation shift factor is the so called PTDF (power transfer distribution factor), describing the situation when the change in generation is not compensated by a reference bus. PTDF determines a change in the power flow at each line when one (1) MW is transferred from one bus of the network to another. We reformulate (2.7) here again as follows: f = y r T B 1 P = H B 1 P = H θ = A P f = A P (2.11) where A is the matrix of generation distribution factor. From the above derivation, we can see that the linear sensitivity vector is equivalent to that of DC power flow. It is very easy to obtain the PTDF via superposition given the direction of the transfer. P T DF = A D = A [0,..., 1,..., 1,...0] T (2.12) where, D is the direction vector (PTDF can also be extended to zonal transfer usage by changing the direction vector to direction matrix) 1 is the injection, and 1 is the withdrawal. 2.3.2.3 Simultaneous Feasibility Test The SFT (Simultaneous Feasibility Test) is widely used in SCED and SCUC [60, 61, 62]. SFT is a Contingency Analysis process. The objective of SFT is to determine violations in all post-contingency states and to produce generic constraints to feed into economic dispatch or unit commitment, where a generic constraint is a transmission constraint formulated according the linear sensitivity coefficients/factors introduced in section 2.3.2.
19 Due to the fast speed and simple application, the SFT procedure is widely used in power industry. The common flowchart is as follows: ED, UC etc. Generic constraints Y SFT Violation? N Finish Figure 2.1: SFT flowchart The ED or UC is first solved without considering network constraints and security constraints. The results are sent to perform the security assessment in a typical power flow. If there is an existing violation, the new constraints are generated using the sensitivity coefficients/factors and are added to the original problem to solve repetitively until no violation exists. 2.3.3 Optimal Power Flow The optimal power flow of OPF has had a long history in its development. It was first discussed by Carpentier in 1962 and took a long time to become a successful algorithm that could be applied in common use [3]. The general form of the OPF is as follows: Min f(x, u, z) (2.13) s.t. g(x, u, z) = 0 h(x, u, z) h max where g are power flow constraints, h are operation constraints. u are decision variables that can be: Generator voltage magnitude & real power Voltage magnitude & angle at slack bus
20 Real power flow through DC lines Phase angles across phase-shifting transformers Turn ratios of tap-changing transformers Admittances of variable reactors and switched capacitor banks Breaker positions by which the network can be reconfigured x are the state (dependent) variables that can be: Real & reactive power at reference bus Reactive power & angle at PV buses Voltage magnitudes & angles at PQ buses z are the exogenous variables (specified) that can be: Real & reactive demands at load buses Tie line flows Admittance matrix The above is the general form, Next we will give the detailed formulation to be used in this work (we will delete the z in the formulation for simplicity purpose). In the following three formulations: f 0 models the cost of preventive control actions, and, for the kth system configuration (k = 0 corresponds to the pre-contingency configuration, while k = 1,..., c correspond to the c post-contingency configurations), x k is the vector of state variables, u k is the vector of preventive and corrective control variables. Equality constraints g k ( ) = 0 is the bus power balance equations and the inequality constraints h k ( ) h max k concerns physical limits of equipments (e.g., bounds on: generators active/reactive powers, transformers equipped with tap-changer ratios, shunt reactance, phase shifters angle, etc.) and operational limits (e.g., limits on branch currents and voltage magnitudes). max k is the
21 vector of maximally allowed variations of control variables between the base case and kth post-contingency state. The formulation of OPF is as follows, Min f 0 (x 0, u 0 ) (2.14) s.t. g 0 (x 0, u 0 ) = 0 h 0 (x 0, u 0 ) h max To check the steady-state security of the power system, security constraints are added to problem (2.14) to form the security-constrained optimal power flow (SCOPF) [59]. Min f 0 (x 0, u 0 ) (2.15) s.t. g k (x k, u 0 ) = 0 k = 0, 1, c h k (x k, u 0 ) h max k k = 0, 1, c In this formulation, all security constraints are presented at the same time. Sometimes the security-constrained optimal power flow also called preventive security-constrained optimal power flow (PSCOPF), and in this dissertation we denote it simply as SCOPF. Compared to the SCOPF, the corrective security-constrained optimal power flow (CSCOPF) was proposed in 1987 [19] and is formulated as follows Min f 0 (x 0, u 0 ) (2.16) s.t. g k (x k, u k ) = 0 k = 0, 1, c h k (x k, u k ) h max u k u 0 max k k = 0, 1, c k = 1, 2, c Compared with SCOPF, the CSCOPF does not require that the system state variables are within limit before and after contingencies. Instead the CSCOPF permits the corrective control to bring the state variables back to the limit after a contingency. In the market environment, the demand bids can also be added to the objective function without changing the structure of the problem.
22 CHAPTER 3 BENDERS DECOMPOSITION Benders decomposition will be used as a principal technique in this dissertation, so a comprehensive introduction of this methodology will be given in this chapter. 3.1 Introduction of the Problem Structure J.F. Benders introduced the Benders decomposition (BD) algorithm for solving large-scale, mixed-integer linear programming problems (MILP), which partition the problem into a programming problem (which may be linear or non-linear, and continuous or integer) and a linear programming problem [9]. A.M. Geoffrion generalized this method, which is generalized Benders decomposition (GBD) and made it applicable to nonlinear problems [10]. Problems for which Benders can be applied are those that have the following structure: Min z = c(x) + d(y) (3.1a) s.t. A(x) b (3.1b) E(x) + F (y) h (3.1c) Constraint (3.1c) is referred to as the coupled constraint and E(x) is called the coupler in this dissertation. In Benders decomposition [9], the second part problem is required to be a linear programming problem, which is convex and dual theory can be applied to. Although in [10] all the objective function and constraints can be completely nonlinear functions, a so called P property is preferred for better performance, which means that the decision variables are partitioned explicitly in the objective function and constraints. Even if the P property holds, the second part problem could be still nonconvex with only weak duality applied, indicating
23 that a dual gap exists. When GBD is mentioned in this dissertation, the P property is applied implicitly. From the literature review (section 1.2), it is observed that most successful applications of generalized Benders decomposition is linear coupled, which means that in addition to the P property the coupler in the coupled constraint is a linear function. We name the linear coupled generalized Benders decomposition as LGBD. LGBD is actually GBD with additional the requirements of P property and linear coupling. Table 3.1: Problem Formulation Summary for BD, GBD and LGBD Num d(y) & F(y) coupler E(x) BD linear linear GBD linear/nonlinear linear/nonlinear LGBD linear/nonlinear linear 3.2 Outline of the Methodology We repeat problem (3.1a) here again and put it into the BD form, which means that all the objective and constraints are linear. For simple explanation, the coupler is put into explicitly linear form. This form is called the standard form in this dissertation. Min z = c(x) + d(y) (3.2a) s.t. A(x) b (3.2b) E x + F (y) h (3.2c) This problem can be decomposed into three subproblems [20]: master problem, feasibility subproblem, and optimality subproblem. Master Problem Decide on a feasible x considering only constraint (3.2b) via what is referred to as the
24 master problem: Min z = c(x) + α(x) (3.3a) s.t. A(x) b (3.3b) where α(x) is a piecewise function of the optimality subproblem optimal value as a function of the master problem decision variable x. z is a lower bound of the whole problem and will be updated iteratively by the optimality subproblem. Feasibility Subproblem In order to check whether (3.2c) is satisfied based on x given in the master problem, a slack vector is introduced and the corresponding subproblem is formulated as: Min ν = 1 T s (3.4a) s.t. F (y) + s h E x (3.4b) Here, 1 T is the vector of ones, and ν > 0 means that violations occur in the subproblem. In order to eliminate the violations, the feasibility cut (3.5) is added to the master problem: ν + λe(x x) 0 (3.5) where λ is the Lagrangian multiplier vector for inequality constraints (3.4b). This problem is called feasibility check subproblem or feasibility problem for short. Optimality Subproblem Decide on a feasible y considering constraint (3.2c) given x from the master problem. ω = Min d(y) (3.6a) s.t. F (y) h E x (3.6b) where ω is the value of α(x) at x. If the solution is not optimal, the optimality cut (3.7) is added to the master problem: ω + πe(x x) α (3.7)
25 where π is the Lagrangian multiplier vector of inequality constraints (3.6b). This subproblem is called the optimality check subproblem or optimality problem because it is used to check the optimality of the master problem according to the Benders Rule. Benders rule: The partition theorem for mixed-variables programming problems [9] provides an important optimality rule on which Benders decomposition is based. The optimal solution (z, x ) from the master problem and the optimal solution y from the optimality subproblem are obtained. If the upper bound z = c(x ) + d(y ) is equal to the lower bound z from the first-stage problem, then the triplet (z, x, y ) is the optimal solution for the complete problem. In other words, the problem is optimal only when its subproblems are also optimal. The process of Benders decomposition is a learning process (try-fail-try-inaccurate-try-- solved) as shown in Fig.3.1. This process is called the Benders process and is described as follows. 1. Solve the master first and then send the optimal solution x to the two subproblems; 2. Check if the optimal solution from the master problem is feasible. If it is not feasible, the feasibility check subproblem will send a feasibility cut back to the master problem so that the master problem can remove this infeasible solution set. 3. Use Benders rule to check if the estimation of the optimality subproblem optimal value from the master problem is accurate enough. If Benders rule is not met, an update estimation of α(x) is sent to the master problem by an optimality cut. If the Benders rule is met, the problem is solved. This process is easily expanded to stochastic programming. The optimal value from the master problem is sent to the multi-scenario optimality problem. The process is exactly the same as in the deterministic case, except that all the optimality cuts and the optimal value from the optimality problem are weighted by the probability of the scenario. The more complicated case would be the scenario tree.
26 master problem (z,x * ) feasibility cut N feasibility subproblem optimality cut Feasible? v(x * )=0? N Y optimality subproblem (z,y * ) Optimal? z-z =0? Y finished Figure 3.1: Benders process A typical procedure using Benders decomposition is illustrated in Fig.3.2. First, a largescale and complex problem, which is hard or impossible to solve as a whole, is studied to decide how to formulate the master problem, feasibility subproblem and optimality subproblem. During this step, the coupled constraints need to be carefully determined. Second, the solution algorithms or solvers for different subproblems should be decided. Although different subproblems can utilize different algorithms or solvers, the requirement is that the algorithms or solvers for the feasibility and optimality subproblem need to provide Lagrange multipliers. Third, the Benders process and Benders rule are used to obtain the final solution. Original Problem Large Complex Master Feasibility Optimality Algorithm Solver Benders Rule Benders Process Solution Benders triplet Small Easy Figure 3.2: Benders procedure
27 3.3 Derivation of the Benders Decomposition Principle The feasibility problem/cut and optimality problem/cut is very meaningful,both mathematically and physically [20] and they will be derived in this section. 3.3.1 Derivation of the Optimality Problem and Cut The optimality cut is used to update the estimation of the function α(x) in problem (3.3a), as reproduced below [20]: Min c(x) + α(x) (3.8) s.t. A(x) b where α(x) is the solution of the optimality problem (3.6a). This problem is reproduced below as (3.9): ω = Min d(y) (3.9) s.t. F (y) h E x The Benders decomposition principle can be derived as follows: the dual of the optimality problem (3.9) can be written as Max π(h E x) (3.10) s.t. πf d Note that the feasible region of problem (3.10), defined by πf d, does not depend on the master problem decision x. The region is a convex polyhedron, and can be characterized by the set of its extreme points (or vertices) Π = { π 1, π 2,, π p} where p is the number of vertices. So we can conclude that the solution of problem (3.10) will be the maximum vertex. At the same time, it will be the optimal solution of problem (3.9) due to duality. Furthermore, we can write the master problem as: Min c(x) + α (3.11) s.t. A(x) b α max {π(h Ex)}
28 We know ω is the optimal solution value of the optimality problem (3.9) for a given master problem solution x. Because the optimal solution values of the primal and dual problems coincide, we can say that ω = π (h Ex ) π h = ω + π Ex. Substituting π h into the expression of the Benders cut π j (h Ex), we obtain ω + π E(x x) α (3.12) The optimality cut (α π i (h Ex)) can be added one at a time to further reduce the size of the master problem by number of the constraints. But it should be noticed that this size reduction will result in multiple iterations. 3.3.2 Derivation of the Feasibility Problem and Cut The solution x from master problem could result in an infeasible solution to the original problem because only a subset of the constraints are considered. With the introduction of the slack variable, the problem (3.4a) will be always feasible. In solving (3.4a), if the objective function is nonzero, implying that the original problem is not feasible, the feasibility cut (ν +λe(x x) 0) is generated. The physical meaning of this cut is as follows: moving x around the x along the direction of the gradient (λe) to make the ν to be zero. For better understanding, the feasibility cut can be put into the form of Newton method (the is changed to =), as x = x + (λe) 1 ν(x ), which is for solving ν(x) = 0. The graph to describe this is shown in Fig.3.3. ν(x) ν=ν(x*) ν(x) Slope, λe x x* x Figure 3.3: Graph description of feasibility cut The above derivation (slack variable approach) is used mostly in power system area and another kind of derivation (extreme ray approach) can be seen in [63]. If (3.9) is infeasible,
29 then the dual (3.10) is unbounded. Then an extreme ray need to be obtained to generate the feasibility cut. Compared with this one, the slack variable derivation we introduced earlier is easier to understand and apply to the practical problem in power system area. Although the derivation is different, the basic purpose is to make the optimality subproblem feasible. 3.3.3 An Example of Benders decomposition The problem is put into the standard form introduced earlier as follows: Min ( 5)x + ( 4)y 1 + ( 3)y 2 (3.13) s.t. ( 1)x + ( 0)y 1 + ( 0)y 2 20 ( 1)x + ( 2)y 1 + ( 3)y 2 12 ( 3)x + ( 2)y 1 + ( 1)y 2 12 where x,y 1, and y 2 are greater or equal to zero. x is an integer. The master problem will be formulated as follows: M in ( 5)x + α (3.14) s.t. ( 1)x + ( 0)y 1 + ( 0)y 2 20 α will be set to a very small number in the first iteration and will be updated in subsequent iterations. The feasibility check problem is formulated as follows: Min ( 0)y 1 + ( 0)y 2 + (+1)s 1 + (+1)s 2 (3.15) s.t. ( 2)y 1 + ( 3)y 2 + (+1)s 1 + (+0)s 2 12 ( 1)x ( 2)y 1 + ( 1)y 2 + (+0)s 1 + (+1)s 2 12 ( 3)x where y 1, y 2, s 1, and s 2 are greater or equal to zero. The optimality check problem is formulated as follows: Min ( 4)y 1 + ( 3)y 2 (3.16) s.t. ( 2)y 1 + ( 3)y 2 12 ( 1)x ( 2)y 1 + ( 1)y 2 12 ( 3)x
30 where y 1 and y 2 are greater than or equal to zero. The dual problem of the optimality check problem is as follows: Max ( 12 + x)π 1 + ( 12 + 3x)π 2 (3.17) s.t. ( 2)π 1 + ( 2)π 2 4 ( 3)π 1 + ( 1)π 2 3 where π 1 and π 2 are greater than or equal to zero. After solving this problem, the following cuts are obtained from different iterations: 4x 24 f easibility cut 3x 12 f easibility cut 9x α 36 optimality cut 2x α 24 optimality cut 5x α 24 optimality cut From the feasibility cut, it is easy to find that x 4. The procedure is further explained in Fig.3.4. The left part is the feasible region of the dual problem (3.17), which will not be affected by the results of the master problem according to (3.10). But when the master problem sends different results, the objective function is different, which causes the optimal value of the optimality problem to jump among all the three candidate vertexes. In the right plot, the estimation of the α as a function of x is plotted according the right plot. After this procedure the master problem becomes a mixed integer programming problem that only depends on x. Due to the non-convexity of the integer problem this problem has multiple solutions. From the optimality cuts information we can obtain the same results as shown in Table.3.2. The underlined values are correponding to the square point in Fig.3.4 The core part of Benders decomposition is to obtain α(x), which is based on dual theory. One brilliant aspect of this algorithm is that the dual problem has a constant feasible region and the optimal value of the optimality problem has to on one of the vertexes of this constant region.
31 π 2 3 2 4 x=0,[-12,-12] x=1,[-11,-9] x=2,[-10,-6] x=3,[-9,-3] x=4,[-8,0] 01 Feasible region of dual problem 0-9 -14 0 1 2 3 4 x 1 3 2-19 -5x -24 α(x) 0.5 1 2 π 1 Figure 3.4: Example of the Benders decomposition Table 3.2: Estimation of the α(x) x 0 1 2 3 4 α 9x 36-36 -27-18 -9 0 α 2x 24-24 -22-20 -18-16 α 5x 24-24 -19-14 -9-4 3.3.4 Some Special Forms of Benders Decomposition The above introduction is the full version of Benders decomposition. Some special forms also exits. No optimality problem. Although there is no y in the objective function of the original problem (3.2a), y still exit in the constraints. This form is very common and possibly one of the most used forms of Benders decomposition, such as SCUC in [47]. No feasibility problem. In some situations, the optimality problem will be always feasible. So the feasibility problem is not necessary. Dual-role feasibility and optimality problem. In some applications, the feasibility and optimality problem can be the same problem. The original form and the above special forms will be used in this dissertation.
32 3.4 Advantages and Limitations of Benders Decomposition Benders decomposition is a very powerful algorithm, with the following advantages: a) De-scaling the problem. For a very large scale problem, the Benders decomposition algorithm can decompose the problem into a certain number of small scale problems, achieving better computational efficiency because computation complexity of most problem is not linear. b) Flexibility. After the optimization problem is decomposed, each subproblem can be solved by any appropriate algorithm/tool. c) Parallel computing. Most of the subproblems of Benders decomposition are independent of each other, and this characteristic can facilitate the possibility of using parallel computing. d) Stochastic programming. For complex stochastic programming problem, Benders decomposition can be a very good solution method. These advantages make Benders decomposition very attractive. But in order to use Benders decomposition, some requirements must be met. I Mathematically the formulation of the original problem must be staircase as shown in the standard form and convexity assumption must be met for the subproblems. The staircase insures that the problem can be decomposed easily and the convexity assumption ensures that the problem converges to a real optimal solution. II Physically the problem can be time-decomposed and/or function-decomposed. Timedecomposition means that the relationship between subproblems are sequential. Functiondecomposition represents a situation in which each subproblem performs different function independently.
33 CHAPTER 4 RISK-BASED OPTIMAL POWER FLOW In the first part of this chapter, a decomposed security-constrained optimal power flow is proposed to show the speed enhancement capability obtained through applying Benders decomposition. In the second part an improved corrective security-constrained optimal power flow is proposed to avoid possible catastrophic consequence of the corrective security-constrained optimal power flow. In the third part, a risk-based optimal power flow is described, which can improve economic benefit while guaranteeing system reliability. 4.1 Decomposed Security-constrained Optimal Power Flow 4.1.1 Introduction Security-constrained optimal power flow (SCOPF) problem is computationally intensive if implemented using a full AC power flow and with numerous contingencies. The computational bottleneck of the SCOPF is due to the scale caused by the number of contingencies. In this work, we target this issue via application of Benders decomposition, to decompose the problem into subproblems associated with each contingency. The proposed formulation is referred to as the decomposed SCOPF, or DSCOPF. 4.1.2 Problem Formulation and Solution Method The formulation of the SCOPF is [59]: Min f 0 (x 0, u 0 ) (4.1a) s.t. g k (x k, u 0 ) = 0 k = 0,, c (4.1b) h k (x k, u 0 ) h max k = 0,, c (4.1c)
34 where f 0 models the cost of control actions, and, for the kth system configuration (k = 0 corresponds to the pre-contingency configuration, while k = 1,, c correspond to the c postcontingency configurations), x k is the vector of state variables, and u 0 is the vector of preventive control variables. In [19], by setting the corrective control to zero, security-constrained optimal power flow is decomposed, but subproblems no longer have OPF structure as shown in (4.1a), and as a result, the numerical tests showed that the computation time is more sensitive to the number of buses of the system than when they have OPF structure. Reference [48] solves the optimal power flow master problem, then minimize the real and reactive power injected by fictitious sources in subproblems, resulting in a very large number of decision variables for large systems. The approach presented in this work is similar to these two approaches in that the master problem is an optimal power flow, but it differs in that the subproblems minimize adjustment of the preventive controls in subproblems. This is advantageous because decision variables in the subproblem include only control variables of the complete problem, reducing computation effort. In addition, the subproblems retain the optimal power flow structure and therefore have similar computation complexity. This section explains why this decomposition increases computational efficiency for the SCOPF and illuminates our approach via three examples. The problem is decomposed for solution via Benders decomposition method and for this problem, the no optimality subproblem special form will be used. In the Benders approach [20] to solving DSCOPF, the master problem finds an optimal solution under normal constraints. Min f 0 (x 0, u 0 ) (4.2) s.t. g 0 (x 0, u 0 ) = 0 h 0 (x 0, u 0 ) h max The c contingency sub-problems are then: Min 1 T ɛ k (4.3) s.t. g k (x 0 k, u 0 + ɛ k ) = 0 h k (x 0 k, u 0 + ɛ k ) h max
35 where 1 T is the vector of ones, ɛ k is the vector representing the preventive control adjustment, and superscript 0 represents that the state variables are different from the normal condition. The purpose of this subproblem is to eliminate these adjustments by passing cuts (constraints) back to the master when the subproblem is infeasible. λ is the Lagrangian multiplier vector of constraints. u 0 is the fixed control from the master problem. In order to eliminate the violations, the feasibility cut (4.4) is added to the master problem [20]: 1 T ɛ k + λ(u 0 u 0 ) 0 (4.4) As shown in Fig.4.1 the optimal power flow (master) is solved first. The resulting control variables are sent to a power flow solver to determine if there is any violation under each credible contingencies. If there is violation, these control variables are sent to the contingency subproblems to generate the Benders cut. All subproblems are independent and can be solved in parallel. The overall problem solution is found when all control provided by the master problem is acceptable for the entire post-contingency situation. Start cut Optimal Power Flow cut Flow Subpro N check Contingency 1 Flow N Subpro check Contingency c Finish Figure 4.1: The flowchart to solve SCOPF parallel with Benders Decomposition 4.1.3 Illustration The approach is illustrated on two 6-bus test systems A6 [49], B6 [3], and on the 24-bus RTS96 [64]. Computing time for three different strategies are given in Fig. 4.2. In all three strategies, the basic OPF solver is the same (run via MATLAB v. R2007a on Dell PC with
36 2.80GHz, 3.75 GB RAM), the contingencies are randomly drawn, and the initial values for nonlinear OPF solver are the same. The parallel case represents an ideal situation which assumes all subproblems can be solved simultaneously without considering communication time. Actual parallel computing times would be slightly more than this ideal case. A solution obtained for a specific number of contingencies using SCOPF is the same as that obtained using DSCOPF in all cases. 12 10 8 SCOPF DSCOPF PARALLEL 15 14 12 10 SCOPF DSCOPF PARALLEL Second 6 4 Second 8 6 4 2 2 0 1 2 3 4 5 6 Number of contingencies (a) Test system A6 0 1 2 3 4 5 6 7 8 9 10 11 Number of contingencies (b) Test system B6 Figure 4.2: Computation time comparison of 6-bus test system We observe that computing time of SCOPF increases quadratically with increase in number of contingencies. In contrast, computing time of DSCOPF increases linearly. The same comparison is performed on RTS 96 with 1, 2, and 3 contingencies, and two typical cases (random contingencies) are shown in Table 4.1, which shows the DSCOPF is faster, accurate and stable than SCOPF under same numerical tolerances; the comparison is consistent with that of the 6-bus systems. We provide a complexity analysis to support these observations. Suppose the OPF problem has k preventive control variables, l state variables, and m constraints. With n contingencies, the SCOPF will have k preventive control variables, l n state variables, and m n constraints. The numerical optimization generally uses the gradient based method which requires computation of the Jacobian. When we calculate the Jacobian, because the number of variables is about 2 n times original OPF problem, the computation time is about n 2 times. So compared to the OPF problem, we estimate computation complexity of
37 the SCOPF as nearly O(n 2 ) and some of the constraints are inequality constraints which make the complexity even worse. In contrast, the complexity of DSCOPF is only O(n), because it only solves a certain number of problems of similar size to the OPF problem. If computed in parallel the complexity of DSCOPF is even less. Figure 4.3 provides insight into the Benders algorithmic process. Case I is line 2-3 outage of the A6 test system, and Case II is line 1-2 outage of the B6 test system. PG6 0.55 0.5 0.45 0.4 0.35 0.3 0.25 Benders cut Solution before cut Solution after cut PG3 1 0.9 0.8 0.7 0.6 Benders cut Solution before cut Solution after cut Generation limit 0.2 0.15 0.1 0.05 Generation limit 0.2 0.4 0.6 0.8 1 PG2 (a) Test system A6 0.5 0.4 0.3 0.3 0.4 0.5 0.6 0.7 0.8 PG2 (b) Test system B6 Figure 4.3: Computation evolution of DSCOPF In case I, unit 1 of the A6 test system is much less costly than the other two, so the OPF will give a result residing at the square point in Fig.4.3a. Then the contingency pushes it to the round point which results in less generation from unit 1. In case II, a similar situation happens and due to network constraints the square point is Table 4.1: Computing speed and results using different strategies of RTS96 # Speed(Second) Results(+30000$) SCOPF DSCOPF PARA. SCOPF DSCOPF i ii i ii i ii i ii i ii 1 79-11 6 11 6 515-512 512 2 2k 17 17 19 13 18 680 676 656 655 3 - - 17 23 13 18 - - 1082 1082
38 not at the minimum of unit 2 as shown in Fig.4.3b. 4.1.4 Conclusion In this section, Benders decomposition is applied to decompose the traditional SCOPF, and the underlying computational complexity is analyzed. This approach results in significantly better computing speed without sacrificing accuracy for both serial and parallel computing strategies. 4.2 Improved Corrective Security-constrained Optimal Power Flow 4.2.1 Introduction CSCOPF extends the feasible solution region of SCOPF by considering the available controllability of the power system. However the system could face voltage collapse and/or cascading overload right after a contingency before the corrective action is taken. So the constraints to avoid these catastrophic situations are applied [65], resulting in what we call the improved CSCOPF (ICSCOPF) here. We use the Benders decomposition method to solve ICSCOPF. 4.2.2 Problem Formulation and Solution Procedure The compact form of ICSCOPF is as follows [65]: Min f 0 (x 0, u 0 ) (4.5) s.t. g k (x k, u k ) = 0 k = 0, 1, c g 0 k (x0 k, u 0) = 0 h k (x k, u k ) h max h 0 k (x0 k, u 0) p k h max u k u 0 max k k = 1, 2, c k = 0, 1, c k = 1, 2, c k = 1, 2, c where p k is a scalar value modeling how much the constraints just after the contingency application are relaxed with respect to the permanent limits.
39 Compared with CSCOPF, for each contingency equality constraints gk 0(x0 k, u 0) = 0 and inequality constraints h 0 k (x0 k, u 0) p k h max impose the existence and viability of the intermediate state reached just after contingency occurrence and before application of corrective control actions. The problem is decomposed as follows. The master problem finds an optimal solution under normal constraints, which is actually an OPF formulation, as follows: Min f 0 (x 0, u 0 ) (4.6) s.t. g 0 (x 0, u 0 ) = 0 h 0 (x 0, u 0 ) h max The controllability check sub-problem is as follows: Min 1 T ɛ k (4.7) s.t. g k (x k, u k ) = 0 k = 1, 2, c h k (x k, u k ) h max u k u 0 ɛ k max k k = 1, 2, c k = 1, 2, c In minimizing the weighted summation of the control slack variable ɛ, if it is zero, it means the corrective control can bring the system back to normal after a contingency happens. The collapse and cascading failure check subproblem is: Min 1 T ϖ k (4.8) s.t. g 0 k (x0 k, u 0 + ϖ k ) = 0 k = 1, 2, c h 0 k (x0 k, u 0 + ϖ k ) p k h max k = 1, 2, c where ϖ k is the vector representing the adjustment of the preventive control. The physical meaning of this subproblem can be explained as follows. Supposing that when the contingency happens, all the controls of the system are frozen. Then the following phenomenon could occur: 1) no severe(cascading overload and voltage collapse) constraint violation; 2) severe low voltage, which could lead to low voltage protection action; 3) voltage
40 collapse, equality condition fails; 4) cascading overload, which could lead to line tripping. If the optimal value of this subproblem is not zero, meaning that the preventive control must be adjusted in some way, situation 2), 3) and/or 4) could happen. The flowchart is shown in Fig. 4.4. Start cut Optimal Power Flow cut N Controllable? Collapse/ Cascading? Y Finish Figure 4.4: Flowchart to solve ICSCOPF with Benders Decomposition 4.2.3 Illustration This approach can be tested using a modified 6-bus test system A6 [49]. In this case, the generator outputs are the only control variables. Voltage magnitude is within [0.90 1.10] (p.u.) and line active power limits are [2, 1, 1, 1, 1, 1, 1] (p.u.). The re-dispatch considers the ramp rates and re-dispatch time to obtain the. For example, if the ramp rate is.5 and corrective time is 1, then = 0.5 1 = 0.5. Case I represents the situation CSCOPF is safe and the re-dispatch effect on voltage collapse is shown in case II, and its effect on cascading overload flow is shown in case III. The dispatch s effect on voltage is not very significant in this case and is not studied here. Case I. No severe constraints violation: The contingency set is branch [1 2 3 4 5 6 7] outage. The post contingency voltage is [0.90 1.10] p.u., and line active power limits are expanded to 1.4 times the original. A 12-minute limit is applied to re-dispatch. No severe constraints are violated in the post-contingency, and thus the CSCOPF solution is feasible.
41 Case II. Voltage collapse: The contingency set is branch [1 2 4 5 6 7] outage. The post contingency voltage is [0.80 1.10] p.u., and line active power limits are expanded by 2.0 times that of original. A 72-minute limit is applied to re-dispatch (This situation is only used to create a voltage collapse and has no practical meaning here). The outage of branch 1 will create a voltage collapse situation which results in a higher cost than the CSCOPF. When this collapse happens, the outputs of generators 2 and 3 are 0.1 and 0.2 respectively. The collapse and cascading failure check subproblem generated a cut which forces the output of generators 2 and 3 to change to 0.1 and 0.2857 respectively, and the voltage collapse is avoided. In this situation, the corrective control time is extended which results in better economic benefit but a risk of voltage collapse. Case III. Cascading overload: The contingency set is branch [1 2 3 4 5 6 7] outage. The post contingency voltage is [0.90 1.10] p.u., and line active power limits are expanded by 1.2 times those of the original. A 24-minute limit is applied to re-dispatch. The outage of the branch 1 will result in too much overload in branch 2 which could cause cascading protection actions. At this time the outputs of generators 2 and 3 are 0.3912 and 0.4197 respectively. The collapse and cascading failure check subproblem generated a cut which forces the output of generators 2 and 3 to change to 0.3952 and 0.4954 respectively, and the cascading overload is avoided. Table 4.2: Results of different strategies CASE SCOPF CSCOPF ICSCOPF I $4652.21 $4443.59 $4443.59 II $4652.21 $3576.80 $3613.63 III $4652.21 $4235.64 $4273.22 The above three cases shows that the CSCOPF can obtain economic benefit but faces the risk of voltage collapse and cascading overload, and the ICSCOPF can avoid these risks. The contingency set is branch [1 2 3 4 5 6 7] outage. In Fig. 4.5, the x-axis is the relaxation rate of the line limit and y-axis is cost in $. From Fig.4.5 we can see that more relaxation rate and longer re-dispatch time results in less cost. But this trend has a marginal effect, that is,
42 Figure 4.5: Sensitivity analysis after some value it will stop improving. 4.2.4 Conclusion In this section, Benders decomposition has been used successfully to solve the ICSCOPF ensuring existence and viability of the short-term equilibrium point using a constant power model after a contingency is applied and before corrective controls may take place. This work could make the CSCOPF not only economically attractive but also practically feasible and it could draw more attention in a market environment which wants to utilize the existing asset more efficiently. 4.3 Risk-based Optimal Power Flow 4.3.1 Introduction Optimal power flow is a fundamental computation in the power industry as an essential tool for market coordinators and owners of multiple generation units. Algorithmic improvements have resulted in significant benefits. To this end, there have been two extremes proposed. The first was the preventive optimal power flow [59], heavily biased towards security. The second was corrective optimal power flow [19] heavily biased towards economy. However, there is concern that the preventive optimal power flow is too costly, and that the corrective optimal power flow is too risky. Although there has been additional work to mitigate this concern
43 by, for example, including voltage collapse and cascading overload constraints in the corrective optimal power flow [65], it is still not clear that the approach maintains an appropriate balance between economy and security. We address this by proposing the risk-based optimal power flow, which is based on the following strategy: if the risk is relatively small, the operation will be in corrective mode; if the risk is relatively large, the dispatch will move toward preventive mode. The risk plays a role to adjust the dispatch according the predicted near-future situation. The approach requires contingency probabilities. All further discussion is referenced to an identified list of contingencies, such that use of the word contingency assumes it is one on the identified list. Normally, the list includes all N-1 contingencies. Traditionally the system operating state is divided into three states: normal state, emergency state and restore state [66]. References [67][68][69] also developed similar classifications with minor differences. In order to explain the relationship of preventive, classical corrective, improved corrective, and risk-based optimal power flow, in this work, these three states are further divided as shown in Figure 4.6. Here, the emergency state is divided into two substates: 1) controllable emergency: If system falls into this sub-state, it can be drawn back to the normal state by the corrective control; 2) uncontrollable emergency: no corrective control can be done to return the system back to normal without load curtailment. The normal state is divided into three sub-states: 1) preventive security state: the system remains within the normal state following occurrence of any contingency; 2) corrective security state: there exists at least one contingency whose occurrence will cause the system to transfer into the controllable emergency state; 3) insecure state: there exists at least one contingency for which occurrence will cause the system to transfer to the uncontrollable emergency state and the system will therefore be forced into a restore state in which load curtailment must take place. The traditional preventive SCOPF [59] is designed to control the system so that it always operates the system in the preventive security state. The corrective SCOPF [19] is designed to improve economic benefit by operating the system in the classical corrective security state (comprised of the corrective security state and part of the insecure state), which exposes the system to the possibility of load curtailment. The improved corrective SCOPF [65] avoids this
44 Normal state Corrective Security state Preventive Security state Controllable Emergency state Restore state Figure 4.6: The system operation state exposure by constraining operation to only the corrective security state (as we have defined it), which eliminates that part of of the classical corrective state that is exposed to the uncontrollable emergency state. Risk-based optimal power flow will use the improved corrective strategy to obtain better economic benefit, subject to additional control on the expected distance of the post contingency operation point to the insecure normal state. This expected distance is a measure of risk. 4.3.2 Modeling of Severity Function Only steady state security is considered in this work, i.e.,no dynamic issue will be considered. So our implementation of risk-based optimal power flow includes overload security (flow violations and cascading overloads) and voltage security (voltage magnitude violations and voltage instability). We view the extreme case of flow violation is cascading overload, and the extreme case of voltage magnitude decrease is voltage collapse. The severity is thus categorized into two types: hard and soft constraints. Hard constraints include cascading overload and voltage collapse, and will not be permitted. Soft constraints include overload and undervoltage; these can be relaxed in some situations. For the two hard-constrained cases, special treatment, which will be described in this work, is performed to eliminate exposure to them. For the soft constraints, the severity is modeled as the distance from the post-contingency operation point to the nearest feasible operating point. This can be represented by several kinds of parameters, such as direct
45 violation, load shedding, and so on. Here we choose the minimum amount of the corrective control (re-dispatch in this section), needed to be applied to return the system to the normal state. 4.3.3 Problem Formulation and Solution Method We first provide the complete problem formulation including all constraints. Then we provide the alternative formulation necessary to apply the proposed Benders decomposition solution method to our problem. 4.3.3.1 Complete Problem Formulation The formulation is as follows, Min f 0 (x 0, u 0 ) + βr(x 0, u 0 ) (4.9a) where β is the weight of risk, f 0 s.t. g k (x k, u k ) = 0 k = 0,, c (4.9b) g 0 k (x0 k, u 0) = 0 k = 1,, c (4.9c) h k (x k, u k ) h max k = 0,, c (4.9d) h 0 k (x0 k, u 0) p k h max k = 1,, c (4.9e) u k u 0 max k k = 1,, c (4.9f) g 1 k (x1 k, u 0 + s k ) = 0 k = 1,, c (4.9g) h 1 k (x1 k, u 0 + s k ) h max k = 1,, c (4.9h) c R(x 0, u 0 ) = P r k s k k = 1,, c (4.9i) k=1 models the cost of preventive control actions, and, for the kth system configuration (k = 0 corresponds to the pre-contingency configuration, while k = 1,..., c corresponds to the c post-contingency configurations), x k is the vector of state variables, u k is the vector of control variables and s k is the severity. Equality constraints (4.9b) and inequality constraints (4.9d) impose the feasibility of the pre-contingency and corrected post-contingency states. On the other hand, equality constraints (4.9c) and inequality constraints (4.9e) impose, for each contingency, the existence and viability of the intermediate
46 state reached just after contingency occurrence and before application of corrective control actions. Equality constraints (4.9b) and (4.9c) are the AC bus power balance equations, while the inequality constraints (4.9d) and (4.9e) account for the physical limits of equipment (e.g., bounds on: generators active/reactive powers, transformers equipped with tap-changer ratio, shunt reactors, phase shifter angles, etc.) and operational limits (e.g., on branch currents and bus voltage magnitudes). Note that p k is a scalar value modeling how much the constraints, just after occurrence of the contingency, are relaxed with respect to the emergency limits. Inequalities (4.9f) are coupling constraints between the base case and post-contingency values of control variables aimed at preventing unrealistic values of corrective control variables; k is the vector of maximal allowed variations of control variables between the base case and kth post-contingency state. Equations (4.9g) and (4.9h) are the severity evaluation constraints, which represent the relief of the severity by re-dispatch. Equation (4.9i) gives the evaluation of the risk index. The superscript (0, 1) represents different set of flow equations, state variables and limit constraints under post contingency short-term situation and post correction situation respectively. 4.3.3.2 Problem formulation in Benders decomposition In this subsection, we describe the application of Benders decomposition to the problem of subsection 4.3.3.1. The master problem finds an optimal solution under normal constraints, which is an ordinary OPF formulation. Min f 0 (x 0, u 0 ) (4.10) s.t. g 0 (x 0, u 0 ) = 0 h 0 (x 0, u 0 ) h max The controllability check subproblem is used to minimize the weighted summation of the control slack variable ɛ. If it is zero, it means the corrective control can bring the system back to the normal state after occurrence of any contingency. The controllability check sub-problem is as
47 follows. Min e T ɛ k (4.11) s.t. g k (x k, u k ) = 0 k = 1, 2, c h k (x k, u k ) h max u k u 0 ɛ k max k k = 1, 2, c k = 1, 2, c where e is the weighting vector, reflecting preference of action. The collapse and cascading failure check subproblem is used to assure the existence and viability of the post-contingency short-term equilibrium and avoid extreme flow violations. Its formulation is as follows: Min a T ϖ k (4.12) s.t. gk 0 (x0 k, u 0 + ϖ k ) = 0 k = 1, 2, c h 0 k (x0 k, u 0 + ϖ k ) p k h max k = 1, 2, c where ϖ k is the vector representing the preventive control adjustment (For generator output, it represented by the combination of positive dummy real injection and positive dummy load), and a is the weighting vector. The physical meaning of this subproblem can be explained as follows. Supposing when the contingency occurs, all the controls of the system are frozen. Then either power is balanced and no severe (within relaxation) constraint violations occur, or one or more of the following occur: 1) severe low voltage, which could lead to operation of low voltage protection; 2) voltage collapse; 3) cascading overload, which could lead to protection action. If the optimal value of this subproblem is not zero, it means the preventive control must be adjusted in some way, otherwise situation 1), 2) and/or 3) could happen. The risk minimization subproblem is c Min R(x 0, u 0 ) = P r k s k k = 1, 2, c (4.13) k=1 s.t. g 1 k (x1 k, u 0 + s k ) = 0 k = 1, 2, c h 1 k (x1 k, u 0 + s k ) h max k = 1, 2, c
48 where s is the vector representing the adjustment of the preventive control, which is the measurement of the severity. Of the above three subproblems, the controllability check (4.11) subproblem and the collapse and cascading failure check subproblem (4.12) are both feasibility check subproblems and as such will generate feasibility cut to return to the master problem. On the other hand, the risk minimization subproblem (4.13) is the optimality check subproblem and will generate optimality cut to return to the master problem. 4.3.3.3 Solution procedure In an earlier section, the original problem is decomposed into one master problem and several subproblems. How these subproblems work together will be described here. The flow chart is shown in Figure 4.7. Start OPF Optimality Cut Controllability check Collapse Cascading Check Feasibility Cut N Feasibility Cut Y N Risk minimization Optimality Check Finish Figure 4.7: Solution procedure of RBOPF using Benders decomposition Step 1. An ordinary OPF problem (4.10) is solved first and a raw guess of risk is made (in the first iteration, the risk guess is zero). Step 2. Check the controllability of the system relative to each credible contingency using (4.11). If there is just one contingency for which it is not possible to use corrective
49 From To R(pu) X(pu) Limit(MW) 1 2 0.005 0.170 200 1 4 0.003 0.258 100 2 3 0.000 0.037 100 2 4 0.007 0.197 100 3 6 0.000 0.018 100 4 5 0.000 0.037 100 5 6 0.002 0.140 100 Table 4.3: Circuit parameters control to bring the system back to the normal state, then a feasibility cut is sent back to the master problem; return to step 1. Step 3. Check if the given dispatch will cause collapse or cascading overload for each credible contingency using (4.12). If any of the defined catastrophic situations can occur, a feasibility cut is sent back to the master problem and return to step 1. Step 4. Minimize the risk using the probability of the contingency and the severity evaluation model introduced earlier in this paper. If the Benders rule is not met, then an optimality cut is sent back to the master problem; return to step 1. Step 5. Output the results. The risk-based optimal power flow is a high level integration of the security-constrained optimal power flow [59], corrective security-constrained optimal power flow [19], improved corrective security-constrained optimal power flow [65] and risk evaluation. 4.3.4 Illustration This approach is illustrated by a simple 6-bus test system A6 modified from [49]. The predefined contingency list consists of N-1 failure of each circuit. The average failure rate is adapted from data characterizing the IEEE reliability test system [64].
50 Bus $ $/MWh $/MW 2 h Ramp(MW/h) 1 176.9 13.5 0.0004 80 2 129.9 32.6 0.001 50 6 137.4 17.6 0.005 20 Table 4.4: Generation unit information Bus G(MW) Q(MVAR) PLoad QLoad 1 [100 220] [-32 68] - - 2 [10 100] [-16 59.5] - - 3 - - 65 27.6 4 - - 65 23.8 5 - - 79.7 25.3 6 [10 50] [-16 42.5] - - Table 4.5: Bus information 4.3.4.1 Case study First, the optimal power flow result is calculated, in which no security constraints are considered, and the cost is $3533.48. Case 1 Security-constrained optimal power flow: In this case, the ordinary security-constrained optimal power flow problem is solved and the cost is $4652.28. Case 2 Security-constrained optimal power flow with post-contingency corrective rescheduling: In this case, the optimal power flow problem is calculated and the post-contingency reschedule has to be finished within a 12-minute interval. The cost is $4443.60. Case 3 Security-constrained optimal power flow with post-contingency corrective rescheduling avoiding collapse and cascading: In this case, the risk minimization part is not integrated and the problem is solved with the emergency operating limit 1.1 times that of the normal thermal limit and operating time for less than 12 minutes. The cost is $4462.41. Case 4 Risk-based optimal power flow
51 In this case, based on Case 3, the risk minimization part is added to assess the possible impact. The cost is $4572.60. In the above four cases, the feasible region of Case 1 corresponds to the preventive security state, the feasible region of Case 2 corresponds to the classical corrective security state, and the feasible region of Case 3 and Case 4 corresponds to the corrective security state. With respect to optimal cost, we can also observe that: Case 1 > Case 4 > Case 3 > Case 2. Case 1 is the most costly because it is the most secure. Case 2 is least costly, but it is exposed to the possibility of collapse or cascading overload. Case 3 is less costly than Case 1 because of the corrective control, and Case 3 is more costly than Case 2 because it avoids the cascading overload. Case 4 is less costly than Case 1 because of the corrective control and Case 4 is more costly than Case 3 because it considers the possible impact. After we solve the above four cases, we can numerically plot the system operation states as shown in Figure 4.8. The preventive security state is plotted according the information from Case 1, and the corrective security state uses the information from Case 3. The classical corrective security state, which does not consider collapse or cascading consistent with the approach of [19], is plotted for comparison from results of Case 2. The generation unit output limit, which is the large rectangle, corresponds to the normal state. In Figure 4.9, the solutions of all the cases are plotted. All uncommented lines in Figure 4.9 have the same meaning as in Figure 4.8. From Table 4.4, we know that unit 1 is the least costly unit and unit 3 is the most costly unit. So when the optimal power flow is solved, the output of units 2 and 3 are at their minimum. In Case 1, the operating point is pushed to the square point by the Benders cuts (security constraints); in Case 2, it is forced to the diamond point by the controllability check subproblem s Benders cuts, which are the classical corrective security constraints; in Case 3, it is moved to the upper triangle point by the collapse and cascading failure check subproblem s Benders cuts (improved corrective security constraints); in Case 4, compared with Case 3, it is further pushed to the right triangle point by the risk minimization subproblem s Benders cuts (risk).
52 PG6 0.6 0.5 0.4 0.3 Corrective security state Classic corrective security state preventive security state 0.2 0.1 Cascading Benders cut Corrective Benders cut Preventive Benders cut Generation limit (normal state) 0 0 0.2 0.4 0.6 0.8 1 PG2 Figure 4.8: The system operation state cut by Benders cut 4.3.4.2 Sensitivity analysis The purpose of the risk index is to function as an indicator which biases decisions conservatively when risk is too high and optimistically when risk is comparatively low, according to the information given. In Figure 4.10, we observe that the solution of risk-based optimal power flow problem will move to security-constrained optimal power flow problem (Case 1) if the contingency probability is too high, and move to the Case 3 solution to obtain better objective value if the contingency probability is too low resulting in a higher security level. The risk under different weight is shown in Figure 4.11. Here, we observe that risk increases under smaller weight. This is reasonable because if we put more weight on the risk, we prefer a more conservative operating pattern and a smaller risk will be obtained. Compared with the deterministic method, the risk-based optimal power flow provides an auto-steering decision-making strategy which can obtain better economic benefit without sac-
53 0.6 0.5 Case 3 Case 2 Case 1 PG6 (pu) 0.4 0.3 Case 4 0.2 0.1 Economic dispatch 0 0 0.2 0.4 0.6 0.8 1 PG2 (pu) Figure 4.9: The solutions of all the cases rificing reliability. 4.3.5 Conclusion In this work, a refined system operation state model is introduced first. The relationship between the sub-states and different optimal power flow strategies is described. Based on this model, the risk-based optimal power flow is proposed and the solution method is given. This approach is illustrated using a test system and the results are satisfying. At the same time the proposed system operation state model is plotted numerically. 4.4 Conclusion The decomposed security-constrained optimal power flow demonstrated the speed enhancement capability of the chosen algorithm: Benders decomposition. The improved corrective security-constrained optimal power flow showed the capability of Benders decomposition to
54 0.56 0.54 PG6 (pu) 0.52 0.5 0.48 0.46 Base case Probability/100 Probability/10 0*Probability 10*Probability SCED Extreme high probability 0.44 0.42 0.45 0.5 0.55 0.6 0.65 PG2 (pu) Figure 4.10: Sensitivity analysis of the risk-based optimal power flow avoid the voltage collapse and cascading overload. Based on these two approaches, risk-based optimal power flow is proposed. This approach can obtain the better economic benefit through reasonable exposure to the risk, and at the same time avoid the catastrophic severity.
55 102 100 98 Risk 96 94 92 90 10*Base weight Base weight 0.1*Base weight 0.01*Base weight Weight Figure 4.11: Sensitivity analysis of the risk-based optimal power flow
56 CHAPTER 5 RISK-BASED UNIT COMMITMENT Economic benefit and reliability are competing objectives in the power industry. Unit commitment, generally performed by the ISO, should identify economically efficient and reliable solutions. In addition, the decision of the unit commitment should also provide a price signal to GENCOs and TRANSCOs to keep their units or network reliable. This chapter reports on the risk-based unit commitment formulation and solution procedure. We assume that the unit commitment is performed for a day-ahead 24 hour period based on supplier offers and system component average failure probabilities. Solution of the problem is obtained via a three-stage stochastic program based on Benders decomposition where unit commitment is performed in stage 1, dispatch is performed in stage 2, and risk minimization is performed in stage 3. 5.1 Problem Formulation and Solving Procedure The problem is formulated using a DC power flow model. The objective function is formulated as follows: Min T N t=1 i=1 st i α it + sd i β it Unit commitment (5.1) +c i p it Economic dispatch + c j=1 P r jt Sev jt Risk evaluation where, T is the period for performing the unit commitment, Nis the number of the units, C is the number of the predefined contingencies, st i is the starting cost of unit i, α it is unit i start at time t, sd i is the shut down cost of unit i, β it means unit i shut down at time t, c i is output cost of unit i, p it is the output of unit i at period t, P r jt is the probability of contingency j at
57 period t, and Sev jt is the severity of contingency j at period t, characterizing load curtailment. Equation (5.1) is written in three lines to show the three stages explicitly, which are unit commitment, economic dispatch and risk evaluation (minimum load curtailment). We will set forth the explicit problem formulation for each stage in the following development. In each case, only the basic constraints are identified, and other constraints can be added as desired without affecting the approach. The first stage problem is as follows: Min T N i=1 n=1 (st iα it + sd i β it ) (5.2) s.t. α it β it = I it I i(t 1) N n=1 P imaxi it D t + R t N n=1 P imini it D t where, I it is the state of unit i at period t, P imax is the maximum output of unit i, P imin is the minimum output of unit i, D t is the load at time t, and R t reservation at time t. The first stage feasibility check problem (minimum load curtailment problem) is as follows, which is run for all time periods. Min d r r (5.3) s.t. p it P imax I it p it P imin I it a 0 (p + r) b 0 a 1 (p + r) b 1 where, r is load curtailment vector, d r is load curtailment penalty factor vector, a 0 is power flow equations and bounds of flows under normal condition, a1 is power flow equations and bounds of flows under contingency condition, and p is generation output vector.
58 The second stage problem is as follows: Min T N i=1 n=1 (c ip it ) (5.4) s.t. p it P imax I it p it P imin I it a 0 (p it ) b 0 The second stage feasibility check problem is as follows. Min d r r + d s s (5.5) s.t. a 1 (p + r) b 1 p p s where, s is the vector of penalty variables for set of coupling constraints, d s is the positive cost vectors, p is the generation output vector given as a known parameter from second stage problem, and is the vector representing the range of the redispatchability of units. The third stage problem is as follows: Min C j=1 P r jt Sev jt (5.6) B s.t. Sev jt = F V a 1 (p + V v) b 1 m=1 V + f v = d p where, V is the load curtailment vector, v is the dummy load needed to balance the whole system, p is the generation injection vector computed by stage 2, f is the network injection vector, d is the bus load vector, F is the penalty factor vector of load curtailment, and B is the bus number. A flow chart of the solution procedure is given in Fig. 5.1 and described as follows: Step 1. A pure unit commitment problem (without considering the network constraint) is solved.
59 N Stage I (UC) Security check Y N Stage 2 (ED) Corrective controllable Y Stage 3 (Min Risk) N N Stage 2&3 Converge Y Stage 1,2&3 Converge Finish Figure 5.1: Flowchart of Solving RBUC using Benders decomposition Step 2. The feasibility check is performed for the normal condition and the predefined contingencies. If infeasible, a Benders cut is generated and returned to step 1. Otherwise, go to step 3. Step 3. Solve a economic dispatch problem without considering contingencies. Step 4. The feasibility check is performed for the predefined contingencies. If infeasible, a Benders cut is generated and returned to step 3. Otherwise, go to next step. Step 5. Solve the risk minimization problem. Step 6. Check if the stage 2 and stage 3 problems converge according the Benders optimal policy. If it has not converged, then return to step 3. If it converges, go to next step. Step 7. Check if the whole problem is converged according the Benders optimal policy. If it is not converged, then return to step 1. If it converges, stop.
60 5.2 Illustration The A6 test system is used here, and a condition monitor is installed in circuit 6. The problem is converged after 4 iteration, as shown in Fig. 5.2. We can see the learning process of Benders decomposition. The lower part is the guessing of the optimal value and the upper is a feasible solution. When the two are equal, that is, when the Benders optimality rule is met, the problem is solved. Figure 5.2: The convergence of the three-stage RBUC In Fig. 5.3, we see that the risk of contingency 1 is much higher then others, because circuit 1 carries most of the inexpensive power from unit 1. So an additional line between bus 1 and bus 2 is suggested. We also can see that in some times the risk is almost zero, and maintenance can be done in these time slots without introducing high risk. We will also see the effect of a real-time condition monitor which can provide a more accurate failure probability. Suppose the monitor installed in circuit 6 provides information that the circuit 6 failure probability should be increased by a factor of 100. From Fig. 5.4a, we see that after obtaining a high failure probability of circuit 6, the flow on circuit 6 decreases. And after the circuit 6 outage, its flow will be picked up by circuit 3. If we use the average failure probability, there will be a violation on circuit 3 after the circuit 6 outage, but after we know a high failure probability of circuit 6, the violation of circuit 3 is avoided or reduced, as shown in Fig. 5.4b.
61 Figure 5.3: Risk of next 24 hours for 7 contingencies (a) flow of circuit 6 before circuit 6 outage (b) flow of circuit 3 after circuit 6 outagee Figure 5.4: The effect of a real-time failure probability 5.3 Conclusion In this chapter, a risk-based unit commitment approach is introduced, which can be used for day-ahead market. And the risk-based economic dispatch, which is a part of RBUC s stage 2 and 3, can also be used for the real-time market. A three-stage stochastic programming approach is proposed to solve the problem. The result is illustrated on a 6-bus test system. The proposed approach is the only illustration for a multiple-stage Benders decomposition application in this dissertation and is very meaningful.